Request-too-large error crashes conversation with no recovery — context and work lost

Bug Description

When a "Request too large (max 20MB)" error occurs, the entire conversation becomes unrecoverable. The error message repeats on every subsequent interaction (including /debug and /feedback), and the user cannot continue from the last successful point. The only option is to manually rewind, losing all work since the last successful message.

Expected Behavior

A transient error like "request too large" should never cause permanent session state loss. Claude Code should:

• Roll back to the state before the failed tool call
• Allow the user to continue from the last successful point
• Preserve the full conversation context accumulated before the error

Steps to Reproduce

1. Have a conversation with significant planning context (~20 minutes of work)
2. Claude Code attempts to read multiple large files that push the request over 20MB
3. "Request too large" error appears
4. Every subsequent message (including /debug, /feedback, typing new prompts) returns the same error
5. Conversation is bricked — user must manually rewind, losing all accumulated context

Actual Behavior

After the 20MB error:

• The error message repeats on every interaction
• /feedback fails with "Error submitting feedback / bug report"
• /debug also triggers the same error
• The conversation is completely unrecoverable
• ~20 minutes of planning work and context is lost
• User must manually rewind to a previous message

Impact

This is a data loss bug. A transient, recoverable error (payload too large) escalates into permanent loss of conversation state and work.

Environment

• Claude Code version: 2.1.42
• OS: macOS (Darwin 25.2.0 arm64)
• Node.js: v25.4.0
• Model: Claude Opus 4.6

Related

Related to #26018 (Read tool sends requests exceeding 20MB API limit without pre-checking payload size).